AI for Fabric and Textile Defect Detection

Textile production loses up to 5% of revenue due to defects missed by manual inspection. A single operator checks 20–30% of rolls at speeds of 20–40 m/min and misses 15–25% of defects due to fatigue. Each roll (50–100 m long) costs 5000–10,000 rubles, and a defective roll shipped to a client is a significant loss. We offer an alternative: an AI computer vision system that inspects 100% of the fabric surface at speeds up to 100 m/min, consistently detecting even microscopic defects. This AI fabric inspection system is based on computer vision for textiles and machine learning methods. It works turnkey: from dataset collection to production line integration.

What problems we solve

Our clients face typical challenges:

• High line speed — at 60 m/min, an operator physically cannot spot a 1 mm hole. Our detector processes each tile in 10 ms on a GPU, providing throughput up to 80+ m/min on an RTX card.
• Defect diversity — holes, stains, broken threads, weaving errors, scratches, folds. We use a two-level architecture: PatchCore for anomaly detection and YOLOv8 for classification. This achieves recall > 0.95 across all classes.
• Fabric variability — smooth, napped, patterned. PatchCore anomaly detection does not require defect labels — only good samples. Retraining for a new fabric type takes one week.

How two-level detection works

Stack: PyTorch, PatchCore, EfficientAD, YOLOv8, OpenCV, TensorRT for inference. The core component is adapting the PatchCore model for line-scan cameras with tile-based processing.

import numpy as np
import cv2
import torch
from anomalib.models import Patchcore, EfficientAD
from ultralytics import YOLO
from dataclasses import dataclass
from typing import Optional

@dataclass
class FabricDefect:
    defect_type: str      # hole / stain / broken_thread / weaving_error / scratch / fold
    severity: str         # minor / major / critical
    bbox: list
    area_px2: int
    confidence: float
    location_pct: tuple   # (x%, y%) - relative position

class FabricDefectDetector:
    """
    Two-level fabric defect detection:
    Level 1: Anomaly detection (PatchCore) — trained only on good fabric
    Level 2: Defect classification (YOLO) — if classification by type is needed

    AITEX Fabric Dataset: 7 defect types, 12 fabric types.
    TILDA: manufacturing defects, 8 classes.
    """
    DEFECT_CLASSES = {
        0: ('hole', 'critical'),
        1: ('stain', 'major'),
        2: ('broken_thread', 'major'),
        3: ('weaving_error', 'major'),
        4: ('scratch', 'minor'),
        5: ('fold', 'minor'),
        6: ('cut', 'critical'),
        7: ('knotting', 'minor')
    }

    def __init__(self, anomaly_model_path: str,
                  defect_model_path: Optional[str] = None,
                  anomaly_threshold: float = 0.5,
                  device: str = 'cuda'):
        self.anomaly_model = Patchcore.load_from_checkpoint(anomaly_model_path)
        self.anomaly_model.eval()
        self.anomaly_threshold = anomaly_threshold

        self.defect_model = YOLO(defect_model_path) if defect_model_path else None
        self.device = device

        # Transformation for tile-based inspection
        from torchvision import transforms
        self.transform = transforms.Compose([
            transforms.Resize((256, 256)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])

    def inspect_fabric_strip(self, strip: np.ndarray,
                               tile_size: int = 256,
                               overlap: int = 32) -> dict:
        """
        Inspect a fabric strip (horizontal frame from line-scan camera).
        Tile-based processing for high-speed lines.
        """
        h, w = strip.shape[:2]
        from PIL import Image

        anomaly_map = np.zeros((h, w), dtype=np.float32)
        count_map = np.zeros((h, w), dtype=np.float32)

        stride = tile_size - overlap

        # Tile extraction
        tiles = []
        tile_positions = []
        for y in range(0, h - tile_size + 1, stride):
            for x in range(0, w - tile_size + 1, stride):
                tile = strip[y:y+tile_size, x:x+tile_size]
                pil_tile = Image.fromarray(cv2.cvtColor(tile, cv2.COLOR_BGR2RGB))
                tensor = self.transform(pil_tile)
                tiles.append(tensor)
                tile_positions.append((x, y))

        if not tiles:
            return {'defects': [], 'anomaly_score': 0, 'pass': True}

        # Batch inference
        batch = torch.stack(tiles)
        with torch.no_grad():
            outputs = self.anomaly_model({'image': batch})
            scores = outputs['pred_score'].cpu().numpy()
            anomaly_maps = outputs.get('anomaly_map')

        # Assemble global anomaly map
        for i, (x, y) in enumerate(tile_positions):
            if anomaly_maps is not None:
                am = anomaly_maps[i].cpu().numpy()
                am_resized = cv2.resize(am, (tile_size, tile_size))
                anomaly_map[y:y+tile_size, x:x+tile_size] += am_resized
                count_map[y:y+tile_size, x:x+tile_size] += 1

        # Normalization
        count_map = np.maximum(count_map, 1)
        anomaly_map /= count_map

        # Detect defective zones
        defects = self._extract_defects(anomaly_map, strip, w, h)

        overall_score = float(np.max(scores))

        return {
            'defects': [d.__dict__ for d in defects],
            'anomaly_score': round(overall_score, 4),
            'anomaly_map': anomaly_map,
            'pass': overall_score < self.anomaly_threshold and len(defects) == 0
        }

    def _extract_defects(self, anomaly_map: np.ndarray,
                          original: np.ndarray,
                          w: int, h: int) -> list[FabricDefect]:
        """Extract defect bboxes from anomaly map"""
        defects = []

        if anomaly_map.max() < 0.3:
            return defects

        # Binarize anomaly map
        norm_map = ((anomaly_map / anomaly_map.max()) * 255).astype(np.uint8)
        _, thresh = cv2.threshold(norm_map, int(self.anomaly_threshold * 255),
                                   255, cv2.THRESH_BINARY)

        # Morphological cleanup
        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
        cleaned = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)
        cleaned = cv2.morphologyEx(cleaned, cv2.MORPH_OPEN, kernel)

        contours, _ = cv2.findContours(cleaned, cv2.RETR_EXTERNAL,
                                         cv2.CHAIN_APPROX_SIMPLE)

        for cnt in contours:
            area = cv2.contourArea(cnt)
            if area < 100:  # too small
                continue

            x, y, bw, bh = cv2.boundingRect(cnt)
            max_anomaly = float(anomaly_map[y:y+bh, x:x+bw].max())

            severity = ('critical' if max_anomaly > 0.85 else
                        'major' if max_anomaly > 0.65 else 'minor')

            # Attempt defect classification
            defect_type = 'unknown'
            if self.defect_model:
                crop = original[y:y+bh, x:x+bw]
                if crop.size > 0:
                    results = self.defect_model(crop, conf=0.35, verbose=False)
                    if results[0].boxes and len(results[0].boxes):
                        cls_id = int(results[0].boxes.cls[0].item())
                        defect_type, severity = self.DEFECT_CLASSES.get(
                            cls_id, ('unknown', severity)
                        )

            defects.append(FabricDefect(
                defect_type=defect_type,
                severity=severity,
                bbox=[x, y, x+bw, y+bh],
                area_px2=int(area),
                confidence=max_anomaly,
                location_pct=(round(x/w*100, 1), round(y/h*100, 1))
            ))

        return defects

Why PatchCore is better than supervised approaches

Supervised models require thousands of labeled defects — expensive and time-consuming. PatchCore uses anomaly detection: trained only on good samples, anomalies are found as deviations from the norm. It stores a representative set of patch-level features from the training images in a memory bank. During inference, it computes the distance between each patch feature and its nearest neighbor in the bank, flagging regions with high distance as anomalies. In practice:

• Collect 500–1000 high-quality fabric images (simply run a roll under a camera).
• Fine-tune on 100–200 frames with defects if classification is needed.
• Saves 5–10x labeling time.

Compare: manual inspection misses 15–25% of defects and works 3–5 times slower. The AI system consistently maintains recall > 95% at speeds up to 100 m/min. That's 2x faster than a human and 4x more accurate. Savings on claims can reach 2 million rubles per year, and a typical system deployment of 1.5–2.5 million rubles pays back in 8–12 months.

Deployment process

1. Production audit — examine line speed, camera types, lighting, available PLCs.
2. Data collection and labeling — capture 10–20 rolls, label all defects.
3. Model training — fine-tune PatchCore + YOLO on your data.
4. Integration — connect cameras, install edge server with GPU, write PLC module.
5. Testing — run 100 rolls, measure recall and precision.
6. Commissioning — train operators, provide documentation and support.

What's included

• Ready model with weights and configuration
• Source code of custom detector with comments
• Real-time monitoring dashboard
• API for integration into your MES/ERP
• Customer team training (3 days)
• 12-month warranty on model performance under unchanged conditions

AI vs manual inspection: comparison

Our engineers have 10+ years of experience in computer vision and have implemented 50+ projects in the textile industry. Savings on salaries and claims pay back deployment in 8–12 months, with a typical ROI of 1.5 million rubles per year. Missing one defective roll is a significant loss that the system prevents in 95% of cases. This fabric quality control system is a key component of textile production automation. Manufacturing AI inspection solutions like ours reduce defects by 30%. Machine vision fabric inspection at high speed ensures consistent quality. This textile quality AI system saves money and improves quality. Request deployment and get first results in 6 weeks. Contact us for a production assessment — we will visit your site and calculate ROI in two weeks. Get a consultation to start the project.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.